//===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
-//
+//
// The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
-//
+//
//===----------------------------------------------------------------------===//
//
// This file promote memory references to be register references. It promotes
//===----------------------------------------------------------------------===//
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
-#include "llvm/Constant.h"
+#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/AliasSetTracker.h"
+#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/StableBasicBlockNumbering.h"
+#include "llvm/Support/Compiler.h"
#include <algorithm>
using namespace llvm;
} else {
return false; // Not a load or store.
}
-
+
return true;
}
namespace {
- struct PromoteMem2Reg {
+ struct VISIBILITY_HIDDEN PromoteMem2Reg {
/// Allocas - The alloca instructions being promoted.
///
std::vector<AllocaInst*> Allocas;
+ SmallVector<AllocaInst*, 16> &RetryList;
DominatorTree &DT;
DominanceFrontier &DF;
const TargetData &TD;
StableBasicBlockNumbering BBNumbers;
public:
- PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
+ PromoteMem2Reg(const std::vector<AllocaInst*> &A,
+ SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt,
DominanceFrontier &df, const TargetData &td,
AliasSetTracker *ast)
- : Allocas(A), DT(dt), DF(df), TD(td), AST(ast) {}
+ : Allocas(A), RetryList(Retry), DT(dt), DF(df), TD(td), AST(ast) {}
void run();
+ /// properlyDominates - Return true if I1 properly dominates I2.
+ ///
+ bool properlyDominates(Instruction *I1, Instruction *I2) const {
+ if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
+ I1 = II->getNormalDest()->begin();
+ return DT[I1->getParent()]->properlyDominates(DT[I2->getParent()]);
+ }
+
+ /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
+ ///
+ bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
+ return DT[BB1]->dominates(DT[BB2]);
+ }
+
private:
void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
std::set<PHINode*> &DeadPHINodes);
- void PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
- void PromoteLocallyUsedAllocas(BasicBlock *BB,
+ bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
+ void PromoteLocallyUsedAllocas(BasicBlock *BB,
const std::vector<AllocaInst*> &AIs);
void RenamePass(BasicBlock *BB, BasicBlock *Pred,
if (AI->use_empty()) {
// If there are no uses of the alloca, just delete it now.
if (AST) AST->deleteValue(AI);
- AI->getParent()->getInstList().erase(AI);
+ AI->eraseFromParent();
// Remove the alloca from the Allocas list, since it has been processed
Allocas[AllocaNum] = Allocas.back();
std::vector<BasicBlock*> DefiningBlocks;
std::vector<BasicBlock*> UsingBlocks;
+ StoreInst *OnlyStore = 0;
BasicBlock *OnlyBlock = 0;
bool OnlyUsedInOneBlock = true;
// Remember the basic blocks which define new values for the alloca
DefiningBlocks.push_back(SI->getParent());
AllocaPointerVal = SI->getOperand(0);
- } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
+ OnlyStore = SI;
+ } else {
+ LoadInst *LI = cast<LoadInst>(User);
// Otherwise it must be a load instruction, keep track of variable reads
UsingBlocks.push_back(LI->getParent());
AllocaPointerVal = LI;
continue;
}
+ // If there is only a single store to this value, replace any loads of
+ // it that are directly dominated by the definition with the value stored.
+ if (DefiningBlocks.size() == 1) {
+ // Be aware of loads before the store.
+ std::set<BasicBlock*> ProcessedBlocks;
+ for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
+ // If the store dominates the block and if we haven't processed it yet,
+ // do so now.
+ if (dominates(OnlyStore->getParent(), UsingBlocks[i]))
+ if (ProcessedBlocks.insert(UsingBlocks[i]).second) {
+ BasicBlock *UseBlock = UsingBlocks[i];
+
+ // If the use and store are in the same block, do a quick scan to
+ // verify that there are no uses before the store.
+ if (UseBlock == OnlyStore->getParent()) {
+ BasicBlock::iterator I = UseBlock->begin();
+ for (; &*I != OnlyStore; ++I) { // scan block for store.
+ if (isa<LoadInst>(I) && I->getOperand(0) == AI)
+ break;
+ }
+ if (&*I != OnlyStore) break; // Do not handle this case.
+ }
+
+ // Otherwise, if this is a different block or if all uses happen
+ // after the store, do a simple linear scan to replace loads with
+ // the stored value.
+ for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end();
+ I != E; ) {
+ if (LoadInst *LI = dyn_cast<LoadInst>(I++)) {
+ if (LI->getOperand(0) == AI) {
+ LI->replaceAllUsesWith(OnlyStore->getOperand(0));
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ LI->eraseFromParent();
+ }
+ }
+ }
+
+ // Finally, remove this block from the UsingBlock set.
+ UsingBlocks[i] = UsingBlocks.back();
+ --i; --e;
+ }
+
+ // Finally, after the scan, check to see if the store is all that is left.
+ if (UsingBlocks.empty()) {
+ // The alloca has been processed, move on.
+ Allocas[AllocaNum] = Allocas.back();
+ Allocas.pop_back();
+ --AllocaNum;
+ continue;
+ }
+ }
+
+
if (AST)
PointerAllocaValues[AllocaNum] = AllocaPointerVal;
// (unspecified) ordering of basic blocks in the dominance frontier,
// which would give PHI nodes non-determinstic subscripts. Fix this by
// processing blocks in order of the occurance in the function.
- for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
- P != PE; ++P)
+ for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
+ PE = S.end(); P != PE; ++P)
DFBlocks.push_back(BBNumbers.getNumber(*P));
// Sort by which the block ordering in the function.
if (AST && isa<PointerType>(PN->getType()))
AST->deleteValue(PN);
- PN->getParent()->getInstList().erase(PN);
+ PN->eraseFromParent();
}
- // Keep the reverse mapping of the 'Allocas' array.
+ // Keep the reverse mapping of the 'Allocas' array.
AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
}
-
+
// Process all allocas which are only used in a single basic block.
for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
- const std::vector<AllocaInst*> &Allocas = I->second;
- assert(!Allocas.empty() && "empty alloca list??");
+ const std::vector<AllocaInst*> &LocAllocas = I->second;
+ assert(!LocAllocas.empty() && "empty alloca list??");
// It's common for there to only be one alloca in the list. Handle it
// efficiently.
- if (Allocas.size() == 1)
- PromoteLocallyUsedAlloca(I->first, Allocas[0]);
- else
- PromoteLocallyUsedAllocas(I->first, Allocas);
+ if (LocAllocas.size() == 1) {
+ // If we can do the quick promotion pass, do so now.
+ if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0]))
+ RetryList.push_back(LocAllocas[0]); // Failed, retry later.
+ } else {
+ // Locally promote anything possible. Note that if this is unable to
+ // promote a particular alloca, it puts the alloca onto the Allocas vector
+ // for global processing.
+ PromoteLocallyUsedAllocas(I->first, LocAllocas);
+ }
}
if (Allocas.empty())
//
std::vector<Value *> Values(Allocas.size());
for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
- Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
+ Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
// Walks all basic blocks in the function performing the SSA rename algorithm
// and inserting the phi nodes we marked as necessary
// The renamer uses the Visited set to avoid infinite loops. Clear it now.
Visited.clear();
- // Remove the allocas themselves from the function...
+ // Remove the allocas themselves from the function.
for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
Instruction *A = Allocas[i];
// Just delete the users now.
//
if (!A->use_empty())
- A->replaceAllUsesWith(Constant::getNullValue(A->getType()));
+ A->replaceAllUsesWith(UndefValue::get(A->getType()));
if (AST) AST->deleteValue(A);
- A->getParent()->getInstList().erase(A);
+ A->eraseFromParent();
}
+
+ // Loop over all of the PHI nodes and see if there are any that we can get
+ // rid of because they merge all of the same incoming values. This can
+ // happen due to undef values coming into the PHI nodes. This process is
+ // iterative, because eliminating one PHI node can cause others to be removed.
+ bool EliminatedAPHI = true;
+ while (EliminatedAPHI) {
+ EliminatedAPHI = false;
+
+ for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
+ NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
+ std::vector<PHINode*> &PNs = I->second;
+ for (unsigned i = 0, e = PNs.size(); i != e; ++i) {
+ if (!PNs[i]) continue;
+
+ // If this PHI node merges one value and/or undefs, get the value.
+ if (Value *V = PNs[i]->hasConstantValue(true)) {
+ if (!isa<Instruction>(V) ||
+ properlyDominates(cast<Instruction>(V), PNs[i])) {
+ if (AST && isa<PointerType>(PNs[i]->getType()))
+ AST->deleteValue(PNs[i]);
+ PNs[i]->replaceAllUsesWith(V);
+ PNs[i]->eraseFromParent();
+ PNs[i] = 0;
+ EliminatedAPHI = true;
+ continue;
+ }
+ }
+ }
+ }
+ }
+
// At this point, the renamer has added entries to PHI nodes for all reachable
// code. Unfortunately, there may be blocks which are not reachable, which
// the renamer hasn't traversed. If this is the case, the PHI nodes may not
// have incoming values for all predecessors. Loop over all PHI nodes we have
- // created, inserting null constants if they are missing any incoming values.
+ // created, inserting undef values if they are missing any incoming values.
//
- for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
+ for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first));
std::vector<PHINode*> &PNs = I->second;
assert(!PNs.empty() && "Empty PHI node list??");
+ PHINode *SomePHI = 0;
+ for (unsigned i = 0, e = PNs.size(); i != e; ++i)
+ if (PNs[i]) {
+ SomePHI = PNs[i];
+ break;
+ }
// Only do work here if there the PHI nodes are missing incoming values. We
// know that all PHI nodes that were inserted in a block will have the same
// number of incoming values, so we can just check any PHI node.
- PHINode *FirstPHI;
- for (unsigned i = 0; (FirstPHI = PNs[i]) == 0; ++i)
- /*empty*/;
-
- if (Preds.size() != FirstPHI->getNumIncomingValues()) {
+ if (SomePHI && Preds.size() != SomePHI->getNumIncomingValues()) {
// Ok, now we know that all of the PHI nodes are missing entries for some
// basic blocks. Start by sorting the incoming predecessors for efficient
// access.
std::sort(Preds.begin(), Preds.end());
- // Now we loop through all BB's which have entries in FirstPHI and remove
+ // Now we loop through all BB's which have entries in SomePHI and remove
// them from the Preds list.
- for (unsigned i = 0, e = FirstPHI->getNumIncomingValues(); i != e; ++i) {
+ for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
// Do a log(n) search of the Preds list for the entry we want.
std::vector<BasicBlock*>::iterator EntIt =
std::lower_bound(Preds.begin(), Preds.end(),
- FirstPHI->getIncomingBlock(i));
- assert(EntIt != Preds.end() && *EntIt == FirstPHI->getIncomingBlock(i)&&
+ SomePHI->getIncomingBlock(i));
+ assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
"PHI node has entry for a block which is not a predecessor!");
// Remove the entry
// entries inserted into every PHI nodes for the block.
for (unsigned i = 0, e = PNs.size(); i != e; ++i)
if (PHINode *PN = PNs[i]) {
- Value *NullVal = Constant::getNullValue(PN->getType());
+ Value *UndefVal = UndefValue::get(PN->getType());
for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
- PN->addIncoming(NullVal, Preds[pred]);
+ PN->addIncoming(UndefVal, Preds[pred]);
}
}
}
/// potentially useless PHI nodes by just performing a single linear pass over
/// the basic block using the Alloca.
///
-void PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
+/// If we cannot promote this alloca (because it is read before it is written),
+/// return true. This is necessary in cases where, due to control flow, the
+/// alloca is potentially undefined on some control flow paths. e.g. code like
+/// this is potentially correct:
+///
+/// for (...) { if (c) { A = undef; undef = B; } }
+///
+/// ... so long as A is not used before undef is set.
+///
+bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
assert(!AI->use_empty() && "There are no uses of the alloca!");
// Handle degenerate cases quickly.
Instruction *U = cast<Instruction>(AI->use_back());
if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
// Must be a load of uninitialized value.
- LI->replaceAllUsesWith(Constant::getNullValue(AI->getAllocatedType()));
+ LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType()));
if (AST && isa<PointerType>(LI->getType()))
AST->deleteValue(LI);
} else {
}
BB->getInstList().erase(U);
} else {
- // Uses of the uninitialized memory location shall get zero...
- Value *CurVal = Constant::getNullValue(AI->getAllocatedType());
-
+ // Uses of the uninitialized memory location shall get undef.
+ Value *CurVal = 0;
+
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
Instruction *Inst = I++;
if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
if (LI->getOperand(0) == AI) {
+ if (!CurVal) return true; // Could not locally promote!
+
// Loads just returns the "current value"...
LI->replaceAllUsesWith(CurVal);
if (AST && isa<PointerType>(LI->getType()))
assert(AI->use_empty() && "Uses of alloca from more than one BB??");
if (AST) AST->deleteValue(AI);
AI->getParent()->getInstList().erase(AI);
+ return false;
}
/// PromoteLocallyUsedAllocas - This method is just like
if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
if (AIt != CurValues.end()) {
- // Loads just returns the "current value"...
- if (AIt->second == 0) // Uninitialized value??
- AIt->second =Constant::getNullValue(AIt->first->getAllocatedType());
- LI->replaceAllUsesWith(AIt->second);
- if (AST && isa<PointerType>(LI->getType()))
- AST->deleteValue(LI);
- BB->getInstList().erase(LI);
+ // If loading an uninitialized value, allow the inter-block case to
+ // handle it. Due to control flow, this might actually be ok.
+ if (AIt->second == 0) { // Use of locally uninitialized value??
+ RetryList.push_back(AI); // Retry elsewhere.
+ CurValues.erase(AIt); // Stop tracking this here.
+ if (CurValues.empty()) return;
+ } else {
+ // Loads just returns the "current value"...
+ LI->replaceAllUsesWith(AIt->second);
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ BB->getInstList().erase(LI);
+ }
}
}
} else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
// don't revisit nodes
if (Visited.count(BB)) return;
-
+
// mark as visited
Visited.insert(BB);
const TargetData &TD, AliasSetTracker *AST) {
// If there is nothing to do, bail out...
if (Allocas.empty()) return;
- PromoteMem2Reg(Allocas, DT, DF, TD, AST).run();
+
+ SmallVector<AllocaInst*, 16> RetryList;
+ PromoteMem2Reg(Allocas, RetryList, DT, DF, TD, AST).run();
+
+ // PromoteMem2Reg may not have been able to promote all of the allocas in one
+ // pass, run it again if needed.
+ std::vector<AllocaInst*> NewAllocas;
+ while (!RetryList.empty()) {
+ // If we need to retry some allocas, this is due to there being no store
+ // before a read in a local block. To counteract this, insert a store of
+ // undef into the alloca right after the alloca itself.
+ for (unsigned i = 0, e = RetryList.size(); i != e; ++i) {
+ BasicBlock::iterator BBI = RetryList[i];
+
+ new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()),
+ RetryList[i], ++BBI);
+ }
+
+ NewAllocas.assign(RetryList.begin(), RetryList.end());
+ RetryList.clear();
+ PromoteMem2Reg(NewAllocas, RetryList, DT, DF, TD, AST).run();
+ NewAllocas.clear();
+ }
}